PacketChatDocs

Use chats, projects, prompts, and knowledge#

PacketChat keeps conversations, reusable prompts, project records, and ingested documents in one installation. Use each surface for the job it currently performs rather than treating every field as automatic model context.

Before you start#

  • Sign in to an active account.
  • Confirm an administrator has enabled at least one provider model.
  • Ask the administrator which projects and knowledge bases are appropriate for the material you will upload.

Start and continue a chat#

  1. Open Chat.
  2. Select a provider model.
  3. Enter the message and send it.
  4. Wait for the streamed response to finish before reloading or navigating away.
  5. Reopen the conversation from the sidebar to continue it.

Conversation history persists on the server. Model parameters or context shown elsewhere in the interface should be verified in the actual request behaviour before relying on them for a controlled workflow.

Branch, attach and export#

Edit and rerun a user message or regenerate the last assistant response to create a sibling branch. Use the branch controls to choose the path to continue; earlier alternatives remain in the saved conversation. These controls apply to ordinary chat; agent-run transcripts retain their chronological history.

Use Attach files before sending an ordinary chat message. PacketChat stores the file and reports whether its text could be extracted. A stored attachment whose context is unavailable has not supplied readable text to the model. Attachments are unavailable in agent chat; use the agent's knowledge bindings instead. Remove an unwanted attachment before sending.

The conversation menu offers Markdown, JSON and text exports. Review the actual downloaded file before sharing it; branch selection alone does not guarantee that other conversation content is excluded.

Share a conversation#

  1. Choose the intended branch and review its content.
  2. Open the conversation menu, select Share, then Create share link.
  3. Copy the link and test it in a signed-out browser.
  4. Reopen Share and revoke the link when access is no longer needed.

Anyone holding an active link can read the shared conversation without an account. A live link is not a fixed snapshot: conversation handling can include messages from other branches as well as content added later. Inspect the actual signed-out link before distributing it, and use a reviewed downloaded export when the recipient must receive fixed text. Revoke the live link before continuing with sensitive material. Revocation blocks later reads; it cannot retract a copy already downloaded.

HTML and SVG artifacts can display a restricted preview; executable artifact content is shown as code. Inspect generated content before using it elsewhere.

Organise projects and prompts#

Projects group the signed-in user's work; reusable prompts are also scoped to their owner. Prompts provide versioned text. Create narrowly named entries, keep sensitive content out of titles, and describe the intended audience in the body.

Selecting or editing a project does not automatically prove that its instruction field entered a model request. Put required instructions into the actual chat or published agent until the calling workflow explicitly supports project context.

Connect a project to PacketAgent#

A project you own can connect to a PacketAgent deployment and show its Worker runs. Prepare these prerequisites in PacketAgent first:

  • A WorkerPackage-bound deployment created through the Packet-product package flow in the intended workspace. A deployment created only in the ordinary Worker workbench cannot be used by this integration.
  • An enabled manual trigger in its immutable Worker version.
  • A workspace-bound PacketChat product credential allowing deployment.inspect and deployment.activate.

See the PacketAgent reference for credential setup and package deployment. Use a separate publisher credential to prepare the package; the project's runtime credential needs only inspection and activation.

  1. Open the project's PacketAgent panel and create a connection.
  2. Enter its name, PacketAgent origin, workspace id, product token and optional default deployment id.
  3. Save the returned notification endpoint and ingest token immediately; the token is shown only at creation or rotation.
  4. Add the returned route configuration to the Worker version's packetchat notification route. Configure the PacketAgent callback origin and secret there, then deploy the updated version.
  5. Start a run from the project and inspect its version, state and run-card updates. Use its open or inspect control to reach the corresponding Worker.

If the deployment lookup fails, check the workspace, package binding and deployment.inspect permission. If activation fails, also check deployment.activate and the enabled manual trigger. PacketChat does not choose an alternative trigger id.

Keep both tokens in their respective secret stores. If the ingest token is lost or exposed, rotate it and update the Worker's notification route before expecting further updates. A project connection does not change the project's instructions into ordinary chat context.

Add knowledge#

  1. Open Knowledge and choose or create a knowledge base.
  2. Upload a supported document within the configured size limit.
  3. Wait for the document status to reach ready.
  4. Search for a phrase that appears only in the uploaded content.

Upload readiness proves ingestion and search. It does not automatically attach the knowledge base to ordinary chat. Bind knowledge to a published agent when that agent should retrieve it.

Verify#

  • A chat response persists after reload.
  • A selected branch survives reload, its downloaded export contains only the content you intend to share, and a revoked share link no longer opens in a signed-out browser.
  • An attached text file reports readable context before its contents are used.
  • A prompt can be reopened and its current version is visible.
  • Project and prompt ownership match the signed-in user.
  • The knowledge document reaches ready and returns a known search result.

If it does not work#

Symptom Action
Send is rejected Re-select an enabled model and ask an administrator to test its account.
Conversation is absent after reload Confirm the response completed and the database is healthy.
Document stays queued Confirm the worker and Redis are running.
Document fails Check worker logs, object storage, file type, and upload limit.
Knowledge search works but chat ignores it Use an agent with the knowledge base explicitly bound.